Anthropic report reveals that China and other countries are abusing Claude for biological and chemical weapons research and military surveillance

Artificial intelligence (AI) developer Anthropic released the latest threat intelligence report on the 10th, describing the AI ​​model Claude abuse cases discovered and blocked in the past eight months. These include Chinese actors using Claude to spy on dissidents, collect public opinions of Taiwanese politicians, and use AI for research that may help develop biological weapons.

5 cases of malicious use discovered by China and Russia blocked

The report lists five cases involving the risk of biological and chemical weapons. The research content includes Quantong virus, "highly pathogenic" avian influenza virus strains, virus families to which smallpox and Mpox (Mpox) belong, as well as venoms and toxins.

Anthropic said that a researcher in an area not served by the company used Claude for several weeks through a "circuit bypass" tool and a virtual private server (VPS) to plan experiments on adapting avian influenza viruses to mammals. Anthropic immediately blocked the account after discovering it and incorporated the relevant cases into the model security protection mechanism.

Anthropic does not support China, Russia and North Korea

Anthropic emphasized that the researchers involved in these cases are still engaged in scientific research and cannot determine whether they have the intention to cause harm, so the identities of the individuals and research institutions have not been disclosed. Reuters reported that Anthropic currently does not support Russia, China, North Korea and other regions.

The report also revealed new categories of AI abuse, including the use of Claude to develop conventional weapons such as guns, missiles, armed drones, and bombs, as well as the locking and control systems for operating these weapons.

Exercise to suppress our radar stations and missile bases

One of the Chinese actors used Claude to design and repeatedly modify a set of Chinese electronic warfare software, which contains 16 modules. It can analyze enemy radars, anti-aircraft missile positions, command posts and communication nodes, evaluate the detection range and electronic interference effects, and then prioritize the suppression order according to the value and weakness of the target. It can also simulate the engagement range of air defense systems such as Patriot and Terminal High Altitude Defense (THAAD).

During the course of the project, the actor changed the simulation scenario to twelve targets in Taiwan, including command bunkers, early warning radar stations, Patriot and "Tiangong" missile sites, major air force bases, and regional combat headquarters. Anthropic assessed that the other party may be a Chinese national defense or military industry researcher and is related to Chinese military research institutions.

Public opinion monitoring of dissidents and politicians

The report also warns that governments and like-minded actors are using AI to increase surveillance of ethnic minorities and dissidents. Anthropic identified and blocked related operations originating from China, Iran, and West Africa.

One of China's "public opinion monitoring" operations used Claude to collect Taiwanese and Western media reports, monitor overseas Chinese expatriates, dissidents, ethnic minorities and Taiwanese politicians, and then grade them according to their political sensitivity and compile them into internal government briefings. Anthropic believes the operation was more likely carried out by contractors working for Chinese government clients.

The report also named China's "Dark Side of the Moon" and "Deep Search", accusing Chinese AI model developers of secretly using Claude to generate answers to respond to their users' questions, and at the same time "distilling" these data to improve their own developed models.